1. Where we use AI
We use AI for one main job: marking students’ written answers and giving feedback. Our marker reads a student’s answer alongside the question, the mark scheme, a model answer and our own marking guidance, and returns a mark broken down by assessment objective with short written feedback.
We do not use AI to:
- mark multiple-choice or short-answer activities — these are checked against a stored answer key;
- write our lessons and activities without human review — every page is written, checked and verified by experienced teachers before it is published;
- make decisions about students such as predicted grades, sets or exam entries; or
- profile students, target advertising or make recommendations to third parties.
Our adaptive practice — which chooses the next question based on how a student is doing — uses fixed, rules-based logic that we have written, not generative AI.
2. What data the AI sees
| Sent to the AI | Never sent to the AI |
|---|---|
| The text of the student’s answer | Name |
| The question, mark scheme and model answer | Email address or username |
| Our marking guidance for that topic | School, class or year group |
| A random reference number | Account or user ID |
We ask students not to put personal information in their answers.
3. Who provides it
We use Anthropic’s Claude models through a business API account. Under Anthropic’s commercial terms:
- our data is not used to train Anthropic’s models;
- request data is deleted within a limited period ([CONFIRM: 30 days by default, or zero data retention if granted]), except where needed to enforce its usage policy or required by law; and
- Anthropic is bound by a data processing agreement and listed on our Sub-processors page.
All AI calls are made from our servers. The AI provider’s keys never reach a student’s browser.
4. Accuracy and teacher control
AI marking is fast and consistent, but it is not perfect. We are open about that:
- AI marks are a learning aid, not an official grade. They are not endorsed by Cambridge International or any awarding body.
- Teachers can review and override any mark. Every override is recorded, and we use these corrections to improve our marking guidance.
- Uncertain marks are flagged. When the marker is not confident, the answer is flagged on the teacher’s dashboard for review.
- We test the marker against answers marked by experienced teachers, and re-test whenever we change the guidance or the model. [CONFIRM: publish the target, e.g. within 1 mark of the teacher’s mark on 90% of answers, once the test set is in place]
- A student’s first attempt is the one that counts. Later attempts are kept as revisions, so the teacher sees the real starting point.
5. Safety and fairness
- Feedback is about the answer, not the student. It quotes the student’s own words and suggests one concrete improvement.
- The marker is instructed to stick to the task and is limited to the subject of the question.
- Feedback length is capped, and there are limits on how often the same answer can be re-marked.
- We monitor for unusual activity and can pause AI marking instantly. If we do, students’ work is saved and marked later.
6. Questions
Schools often want to know more before adopting AI tools, and we welcome the questions. Email privacy@secondnaturelearning.com.
Questions about this document? Email legal@secondnaturelearning.com. Back to the Legal hub